Exploring the use of machine learning for the assessment of skeletal fracture morphology and differentiation between impact mechanisms: A pilot study. Issue 2 (28th January 2022)

• Record Type: Journal Article
• Authors: Dempsey, Nicholas; Bassed, Richard; Amarasiri, Rasika; Blau, Soren
• Is Part Of: Journal of forensic sciences. Volume 67:Issue 2(2022)
• Page Start: 683
• Page End: 696
• Publication Date: 2022-01-28
• Subjects: convolutional neural networks -- deep learning -- forensic anthropology -- image classification -- machine learning -- skeletal trauma; Medical jurisprudence -- Periodicals; Forensic sciences -- Periodicals; Forensic Medicine -- Periodicals; Gerechtelijke geneeskunde; Gerechtelijke chemie; Gerechtelijke psychiatrie; 363.2505
• DOI: 10.1111/1556-4029.14996
• Languages: English
• ISSNs: 0022-1198

Abstract

Abstract: Analyzing and interpreting traumatic injuries is a fundamental aspect of routine forensic case work. As the human skeleton can be impacted through a combination of loading mechanisms and varying impact energies, the analysis and interpretation of skeletal trauma can be complex. Therefore, it is imperative that the reliability of techniques used for analysis are well‐established. There is growing interest in machine learning (ML) in medicine (especially radiology) regarding the use of image classification (a subset of ML) to categorize and predict classes of medical images. Therefore, the feasibility of using image classification for skeletal trauma analysis should be explored for its benefits to forensic pathology and anthropology. The method explored in this paper examined the potential for machine learning, using three dimensional (3D) convolutional neural networks (CNNs), to assess whether morphological features of skeletal trauma to the femur can be used to differentiate between impact mechanisms within a forensic population. The objective of this study was to assess if morphological differences in femoral fractures seen in post‐mortem‐computed tomographic images (PMCT) could be categorized according to mechanism, specifically horizontal impacts resulting from pedestrian motor vehicle impacts (PMVIs) and vertical impact s resulting from high impact falls. Final model results indicated an accuracy between 69.95%–72.86% and 63.08%–66.24% validation. Although these results mean the method could not be practically used in its current form, as a proof of concept, there is potential for it to be developed as a tool to assist in classifying complex fracture states.